Intention recognition method and device, medium and electronic equipment

Abstract

The invention belongs to the technical field of computers, and relates to an intention recognition method and device, a medium and electronic equipment. The method comprises the steps of obtaining a dialogue statement, and performing model preprocessing on the dialogue statement by using a deep learning model to obtain a target statement; obtaining an intention filling template, and performing semantic recognition processing on the target statement by using the language model to obtain semantic information; and performing intention filling processing on the semantic information according to the intention filling template to obtain a target intention. According to the method, the problems of low data quality caused by wrong characters, missing characters and the like of the dialogue statements and high translation cost of multilingual statements are solved, a universal method capable of performing intention recognition on the multilingual dialogue statements is provided, the intention recognition accuracy is improved from multiple angles, the user experience is greatly optimized, and the intent recognition efficiency is improved. And the user backflow degree is improved to a certain extent.

Description

Technical Field

[0001] The present disclosure belongs to the field of computer technology, and particularly relates to an intention recognition method, an intention recognition device, a computer-readable storage medium, and an electronic device. Background Art

[0002] With the development of the Internet, online shopping has gradually become popular among thousands of households. When shopping, users often need to communicate and interact with merchants. The emergence of intelligent customer service is to relieve the pressure on merchants, identify the intentions of some problems feedback by users, and then perform corresponding operations. Therefore, intention recognition has always been the most important part and the most challenging topic in the intelligent customer service system. Currently, mainstream e-commerce companies mainly establish the relationship between user problems and actual intentions through manual labeling, and generate systems such as knowledge bases or knowledge graphs, and then train algorithms through a large number of data labels so that the algorithms can identify the intentions of users.

[0003] However, due to problems such as missing or misspelled words that are extremely likely to occur in user conversations, the data quality of user conversations is low. Moreover, the scale of the proprietary data set is small, resulting in poor performance of the algorithm in the intention recognition task. Not to mention that different models need to be trained for independent data sets collected in different languages, which further exacerbates the problem of too small data scale and also reflects the poor generality of existing algorithms. In addition, the small number of parameters of existing algorithms also affects the accuracy of intention recognition to a certain extent. Summary of the Invention

[0004] To overcome the problems in the related art, the present disclosure provides an intention recognition method, an intention recognition device, a computer-readable storage medium, and an electronic device.

[0005] According to the first aspect of the embodiments of the present disclosure, an intention recognition method is provided, and the method includes:

[0006] Obtain a dialogue statement, and use a deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement;

[0007] Obtain an intention filling template, and use a language model to perform semantic recognition processing on the target statement to obtain semantic information;

[0008] Perform intention filling processing on the semantic information according to the intention filling template to obtain a target intention.

[0009] Optionally, the using a deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement includes:

[0010] Call the programming interface of the deep learning model to perform statement correction processing on the dialogue statement to obtain a target statement; and / or

[0011] Call the programming interface of the deep learning model to perform statement translation processing on the dialogue statement to obtain a target statement; and / or

[0012] Call the programming interface of the deep learning model to perform sensitive word filtering on the dialogue statement to obtain a target statement.

[0013] Optionally, the using the deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement includes:

[0014] Obtain prompt information;

[0015] According to the prompt information, call the programming interface of the deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement.

[0016] Optionally, the according to the prompt information, calling the programming interface of the deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement includes:

[0017] According to the prompt information, call the programming interface of the deep learning model to perform statement correction processing on the dialogue statement to obtain a target statement; and / or

[0018] According to the prompt information, call the programming interface of the deep learning model to perform statement translation processing on the dialogue statement to obtain a target statement; and / or

[0019] According to the prompt information, call the programming interface of the deep learning model to perform sensitive word filtering on the dialogue statement to obtain a target statement.

[0020] Optionally, the deep learning model includes: a generative pre-trained sequence model based on the attention mechanism.

[0021] Optionally, the using the language model to perform semantic recognition processing on the target statement to obtain semantic information includes:

[0022] Input the target statement into the fine-tuned language model, so that the fine-tuned language model performs semantic recognition processing on the target statement to obtain semantic information.

[0023] Optionally, before the inputting the target statement into the fine-tuned language model, the method further includes:

[0024] Obtain original sample data, and perform sample preprocessing on the original sample data to obtain target sample data;

[0025] Based on the target sample data, use the supervised fine-tuning technique to perform model tuning processing on the pre-fine-tuned language model to obtain a fine-tuned language model.

[0026] Optionally, before inputting the target statement into the fine-tuned language model, the method further includes:

[0027] Obtain original sample data, and perform sample preprocessing on the original sample data to obtain target sample data;

[0028] Based on the target sample data, use reward modeling technology to perform model tuning on the language model before fine-tuning to obtain a fine-tuned language model.

[0029] Optionally, before inputting the target statement into the fine-tuned language model, the method further includes:

[0030] Obtain original sample data, and perform sample preprocessing on the original sample data to obtain target sample data;

[0031] Based on the target sample data, use human feedback reinforcement learning technology to perform model tuning on the language model before fine-tuning to obtain a fine-tuned language model.

[0032] Optionally, the performing sample preprocessing on the original sample data to obtain target sample data includes:

[0033] Perform data cleaning on the original sample data to obtain first sample data;

[0034] Perform statement preprocessing on the first sample data to obtain second sample data;

[0035] Perform data augmentation on the second sample data to obtain target sample data.

[0036] Optionally, the intention filling template includes: dialogue questions, dialogue demands, and dialogue entities.

[0037] Optionally, the performing statement preprocessing on the first sample data to obtain second sample data includes:

[0038] Perform model preprocessing on the first sample data to obtain third sample data;

[0039] Perform label partitioning on the third sample data according to the intention filling template to obtain second sample data.

[0040] Optionally, the performing intention filling processing on the semantic information according to the intention filling template to obtain the target intention includes:

[0041] Perform intention filling processing on the semantic information according to the intention filling template to obtain a category intention;

[0042] Determine the target intention according to the filling situation of the category intention in the intention filling template.

[0043] Optionally, determining the target intent according to the filling situation of the category intent in the intent filling template includes:

[0044] When the intent filling template fills the category intent, determining the category intent in the intent filling template as the target intent; or

[0045] When the intent filling template does not fill the category intent, obtaining the next statement of the dialogue statement to determine the target intent according to the dialogue statement and the next statement.

[0046] According to a second aspect of the embodiments of the present disclosure, there is provided an intent recognition device, including:

[0047] A statement acquisition module, configured to acquire a dialogue statement and perform model preprocessing on the dialogue statement by using a deep learning model to obtain a target statement;

[0048] A semantic recognition module, configured to acquire an intent filling template and perform semantic recognition processing on the target statement by using a language model to obtain semantic information;

[0049] An intent filling module, configured to perform intent filling processing on the semantic information according to the intent filling template to obtain a target intent.

[0050] According to a third aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the program instructions are executed by a processor, the steps of the intent recognition method provided in any one of the first aspects of the present disclosure are implemented.

[0051] According to a fourth aspect of the embodiments of the present disclosure, there is provided an electronic device, including:

[0052] A processor;

[0053] A memory for storing executable instructions of the processor;

[0054] Wherein, the processor is configured to: execute the executable instructions to implement the steps of the intent recognition method provided in any one of the first aspects of the present disclosure.

[0055] The technical solutions provided by the embodiments of the present disclosure may include the following beneficial effects:

[0056] In the method and device provided by the exemplary embodiments of the present disclosure, a deep learning model is used to perform model preprocessing on dialogue statements, constructing high-quality target statements, solving the problems of low data quality caused by misspelled words and missing words in dialogue statements, as well as the high cost of multilingual statement translation. Further, it provides data support for providing a general intention recognition method, improving the accuracy of intention recognition from the data perspective, and also being able to enhance the stability of the method by filtering sensitive words or content. Further, a language model is used to perform semantic recognition processing on the target statements, leveraging the powerful reasoning ability of the language model and the learning advantages of more hyperparameter factors for the intention recognition task, improving the accuracy of intention recognition from the model itself. In addition, through the language model, it is also possible to effectively expand and enhance the dataset, forming a large-scale and high-quality dataset for the intention recognition task, and being able to deploy the language model locally to ensure the privacy and security of the data, which has important significance in practical applications. Furthermore, filling the semantic information into the corresponding intention filling template to obtain the user's target intention provides a basis and foundation for the timely execution of business logic, providing a general method for intention recognition of multilingual dialogue statements, optimizing the accuracy of intention recognition from multiple perspectives, greatly optimizing the user experience, and improving the user return rate to a certain extent.

[0057] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present disclosure and used together with the specification to explain the principles of the present disclosure.

[0059] Figure 1 Schematically shows a flowchart of an intention recognition method in an exemplary embodiment of the present disclosure;

[0060] Figure 2 Schematically shows a flowchart of a method for model preprocessing in an exemplary embodiment of the present disclosure;

[0061] Figure 3 Schematically shows a flowchart of another method for model preprocessing in an exemplary embodiment of the present disclosure;

[0062] Figure 4 Schematically shows a flowchart of a method for further model preprocessing in an exemplary embodiment of the present disclosure;

[0063] Figure 5 Schematically shows a flowchart of a method for model tuning processing in an exemplary embodiment of the present disclosure;

[0064] Figure 6 Schematically shows a flowchart of a method for sample preprocessing in an exemplary embodiment of the present disclosure;

[0065] Figure 7 Schematically shows a flowchart of a method for statement preprocessing in an exemplary embodiment of the present disclosure;

[0066] Figure 8 Schematically shows a flowchart of a method for another model tuning process in an exemplary embodiment of the present disclosure;

[0067] Figure 9 Schematically shows a flowchart of a method for yet another model tuning process in an exemplary embodiment of the present disclosure;

[0068] Figure 10 Schematically shows a flowchart of a method for intent filling process in an exemplary embodiment of the present disclosure;

[0069] Figure 11 Schematically shows a flowchart of a method for determining a target intent in an exemplary embodiment of the present disclosure;

[0070] Figure 12 Schematically shows a flowchart of an intent recognition method in an application scenario in an exemplary embodiment of the present disclosure;

[0071] Figure 13 Schematically shows a flowchart of a method for sample preprocessing in an application scenario in an exemplary embodiment of the present disclosure;

[0072] Figure 14 Schematically shows a flowchart of a method for model tuning process in an application scenario in an exemplary embodiment of the present disclosure;

[0073] Figure 15 Schematically shows a structural diagram of an intent recognition device in an exemplary embodiment of the present disclosure;

[0074] Figure 16 Schematically shows a structural diagram of another intent recognition device in an exemplary embodiment of the present disclosure;

[0075] Figure 17 Schematically shows a structural diagram of yet another intent recognition device in an exemplary embodiment of the present disclosure. Detailed implementation manners

[0076] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0077] It should be noted that all actions of obtaining signals, information, or data in the present disclosure are carried out on the premise of complying with the corresponding data protection regulations and policies of the country where the location is located and with the authorization given by the owner of the corresponding device.

[0078] With the development of the Internet, online shopping has gradually become popular among thousands of households. When shopping, users often need to communicate and interact with merchants. The emergence of intelligent customer service is to relieve the pressure on merchants, identify the intentions of some problems feedback by users, and then perform corresponding operations.

[0079] Intention recognition, as the most important part of the intelligent customer service system, has always been a challenging topic.

[0080] Currently, mainstream e-commerce companies mainly establish the relationship between user problems and actual intentions through manual labeling, and generate systems such as knowledge bases or knowledge graphs, and then train algorithms through a large number of data labels so that the algorithms can identify the intentions of users.

[0081] Mainstream e-commerce companies have constructed a Chinese dialogue corpus dataset through their own data within China, and constructed two dialogue model algorithms. The proposed algorithms have also verified the effectiveness of intention recognition on the dataset.

[0082] Specifically, e-commerce companies have constructed a corpus (dataset) of multi-turn conversations using domestic data. This corpus contains more than 1 million after-sales exchanges between users and customer service personnel in the e-commerce scenario. After the dataset is established, e-commerce companies have proposed two algorithms improved based on retrieval values and conducted training. The results show that the proposed algorithms have a significant recognition effect on the intentions of multi-round conversations.

[0083] However, since users are very likely to have typos, misspelled words, etc. when chatting, there may be a large number of problems such as word order or typos in the original dataset.

[0084] When there are multiple typos in the sentences input by users when chatting, it is relatively easy for people to understand, but it is relatively difficult to identify when used as data to train algorithms, which is likely to result in invalid or dirty data. These abnormal sentences are very likely to cause problems such as incorrect intention recognition or deviation in intention recognition.

[0085] The quality of the intent recognition algorithm is not only related to the quality of the dataset, but also closely related to the quantity of the dataset. In the era of large models, the more and higher-quality data can train a higher-quality model. Although the algorithm trained by an e-commerce company using 1 million pieces of data as the dataset has relatively good results, this amount of data is still too small as the scale of the proprietary dataset to ensure the effect of the intent recognition task.

[0086] In addition, for the situation of multiple languages on foreign platforms, currently, mainly customer service in the corresponding language is hired for translation and reply. When training corresponding intent recognition models for different languages, independent datasets may be required for training, and even different algorithm models need to be trained for different languages according to the characteristics of the languages. This method will not only lead to inaccurate translation of multi-language sentences, resulting in too high a cost of using artificial customer service for translation, but also lead to the problem of too poor generality of intent recognition, and the training and maintenance of the model are also very complicated and time-consuming.

[0087] The deep learning algorithms currently trained and adopted are usually models such as the LSTM (Long Short-Term Memory) model. These algorithms have fewer parameters, and the models themselves have deficiencies in the intent recognition task, resulting in less than ideal accuracy of intent recognition.

[0088] In view of the problems existing in the related technologies, the present disclosure provides an intent recognition method. Figure 1 It is a flowchart of an intent recognition method shown according to an exemplary embodiment, as Figure 1 shown. The method may at least include the following steps:

[0089] Step S110. Obtain a dialogue sentence, and perform model preprocessing on the dialogue sentence using a deep learning model to obtain a target sentence.

[0090] Step S120. Obtain an intent filling template, and perform semantic recognition processing on the target sentence using a language model to obtain semantic information.

[0091] Step S130. Perform intent filling processing on the semantic information according to the intent filling template to obtain a target intent.

[0092] In an exemplary embodiment of the present disclosure, a deep learning model is used to perform model preprocessing on dialogue statements, constructing high-quality target statements, solving the problems of low data quality caused by misspelled words and missing words in dialogue statements, as well as the high cost of multilingual statement translation. Further, it provides data support for providing a general intention recognition method, improving the accuracy of intention recognition from a data perspective, and also being able to enhance the stability of the method by filtering sensitive words or content. Further, a language model is used to perform semantic recognition processing on the target statements, leveraging the powerful reasoning ability of the language model and the learning advantages of more hyperparameter factors for the intention recognition task, improving the accuracy of intention recognition from the model itself. In addition, through the language model, it is also possible to effectively expand and enhance the dataset, forming a large-scale and high-quality dataset for the intention recognition task, and being able to deploy the language model locally to ensure the privacy and security of the data, which is of great significance in practical applications. Furthermore, filling the semantic information into the corresponding intention filling template to obtain the user's target intention provides a basis and foundation for the timely execution of business logic, providing a general method for intention recognition of multilingual dialogue statements, optimizing the accuracy of intention recognition from multiple perspectives, greatly optimizing the user experience, and to a certain extent enhancing the user return rate.

[0093] The following will elaborate on each step of the intention recognition method in detail.

[0094] In step S110, a dialogue statement is obtained, and the deep learning model is used to perform model preprocessing on the dialogue statement to obtain a target statement.

[0095] In an exemplary embodiment of the present disclosure, when a user enters the customer service dialogue platform, they can send a dialogue statement to the merchant; the dialogue statement can also be content sent by developers during the development process to guide subsequent processing. This exemplary embodiment does not make any special limitations on this.

[0096] After obtaining the dialogue statement, the deep learning model can be used to perform model preprocessing on the dialogue statement.

[0097] In an optional embodiment, the deep learning model includes: a generative pre-trained sequence model based on the attention mechanism.

[0098] Specifically, GPT (Generative Pre-trained Transformer) is a language generation model based on Transformer (a sequence model based on the attention mechanism). It learns language patterns and contexts by pre-training on large-scale text data and can then be used to generate coherent text with context understanding ability.

[0099] Chat-GPT (Chat Generative Pre-trained Transformer, a chatbot program) is a version that is fine-tuned and refined based on the GPT model to make it more suitable for dialogue generation tasks. Chat-GPT has been specifically optimized and trained in specific domains to provide more accurate, relevant, and creative dialogue responses.

[0100] Therefore, Chat-GPT can be further utilized to implement the process of model preprocessing.

[0101] In an alternative embodiment, Figure 2 A schematic flowchart of a method for model preprocessing is shown, as Figure 2 shown. The method may at least include the following steps: In step S210, the programming interface of the deep learning model is called to perform statement correction processing on the dialogue statement to obtain a target statement.

[0102] When the deep learning model is Chat-GPT, the API (Application Programming Interface) of Chat-GPT can be called, and this API interface is used to perform statement correction processing on the dialogue statement to obtain the corresponding target statement.

[0103] In step S220, the programming interface of the deep learning model is called to perform statement translation processing on the dialogue statement to obtain a target statement.

[0104] When the deep learning model is Chat-GPT, the API interface of Chat-GPT can be called, and this API interface is used to perform statement translation processing on the dialogue statement to obtain the corresponding target statement.

[0105] In step S230, the programming interface of the deep learning model is called to perform sensitive word filtering on the dialogue statement to obtain a target statement.

[0106] When the deep learning model is Chat-GPT, the API interface of Chat-GPT can be called, and this API interface is used to perform sensitive word filtering on the dialogue statement to obtain the corresponding target statement.

[0107] In this exemplary embodiment, during the user's use, the model preprocessing of the dialogue statement can be achieved by calling the programming interface of the deep learning model, which solves the problem of incorrect intention recognition or deviation caused by abnormal statements such as typos, missing words, or sensitive words in the dialogue statement, improves the data quality in the intention recognition process, and also provides a solution to the problem of high costs caused by inaccurate translation of multi-language statements using manual translation.

[0108] In an alternative embodiment, Figure 3 FIG. shows a schematic flow diagram of another method for model preprocessing, as Figure 3 shown. The method may at least include the following steps: In step S310, prompt information is obtained.

[0109] When a developer inputs a dialogue statement during the development process, in order to be able to call the API interface of Chat-GPT for model preprocessing, a Prompt can be written using prompt engineering to guide models such as Chat-GPT through this Prompt.

[0110] Simply put, a Prompt is an instruction given by a developer to an AI (Artificial Intelligence). A Prompt can be a piece of text, such as the text during a conversation with Chat-GPT, or a parameter description in a certain format, such as software for AI drawing.

[0111] Corresponding to the Prompt is Prompt Engineering (PE). PE is a concept in the field of artificial intelligence, especially in the field of Natural Language Processing (NLP). PE usually processes by converting a problem into a specific format of input and using predefined templates, rules, and algorithms, enabling the AI to better understand the task and give corresponding answers.

[0112] The advantage of PE is that it can make the AI understand the task more flexibly and precisely, and can reduce misunderstandings and errors caused by unclear language expressions, enabling it to accurately and reliably execute specific tasks.

[0113] When interacting with the AI, the basic principles of the Prompt are very important. When submitting a Prompt, follow the following principles to ensure that the machine can better understand the user's intention and give corresponding answers:

[0114] (1) Clear task description: When submitting a Prompt, describe the specific information of the task as clearly and explicitly as possible, including the task objective, required operations, relevant conditions, etc.

[0115] (2) Use common vocabulary: When submitting a Prompt, use common vocabulary and language expressions, and avoid using rare words and complex sentence patterns so that the machine can understand more easily.

[0116] (3) Consider the context and context: When submitting a Prompt, consider the context and context environment in which it is located so that the machine can obtain more information from the language environment to understand the user's intention.

[0117] (4) Provide diverse information: When submitting a Prompt, try to provide diverse information, including text, images, voice, etc., to facilitate the machine's more comprehensive understanding of the user's needs.

[0118] (5) Determine the response form: When submitting a Prompt, determine the response form, such as text, voice, etc., and ensure that the machine can reasonably parse and output the response information.

[0119] Further principles can include providing the machine with sufficient context information, organizing the Prompt clearly structured to make it easier for the machine to process, using simple and clear questions to guide the user to express suggestions, so as to better understand the needs.

[0120] In addition, during the interaction, the user can also try to ask questions in short sentences so that the AI robot can answer questions more quickly and accurately.

[0121] The writing patterns of Prompts can include Specific Instructions, Instruction Template, By Proxy, and By Demonstration.

[0122] Among them, in the Specific Instructions mode, some specific information, such as questions or keywords, is provided to the model, and the model needs to generate text related to this information. This mode is usually used to generate answers, explanations, recommendations, etc. The specific information can be a single question or multiple keywords, depending on the requirements of the task.

[0123] In the Instruction Template mode, some clear instructions are provided to the model, and the model can generate text based on these instructions. This mode is usually used to generate texts that require clear instructions, such as technical specifications, operation manuals, etc. The instructions can be a single sentence or multiple paragraphs, depending on the requirements of the task.

[0124] In the By Proxy mode, the user can ask Chat-GPT to generate a response in a specific identity, role, or by impersonating a specific person, role, or object. This mode is usually used to simulate the language style and context of a specific person and generate conversations, responses, or other forms of text in a specific scenario.

[0125] In the By Demonstration mode, some example texts are provided to the model, and the model needs to generate text similar to the example texts. This mode is usually used to generate texts similar to the given examples, such as automatically generating emails, product descriptions, news reports, etc. The example texts can be a single sentence or multiple paragraphs, depending on the requirements of the task.

[0126] In step S320, according to the prompt information, the programming interface of the deep learning model is called to perform model preprocessing on the dialogue statement to obtain the target statement.

[0127] In an alternative embodiment, Figure 4 The flowchart of the method for further performing model preprocessing is shown, as Figure 4 shown, the method may at least include the following steps: In step S410, according to the prompt information, the programming interface of the deep learning model is called to perform statement correction processing on the dialogue statement to obtain the target statement.

[0128] In step S420, according to the prompt information, the programming interface of the deep learning model is called to perform statement translation processing on the dialogue statement to obtain the target statement.

[0129] Considering the problem of foreign users using multiple languages, the prompt information can be used to guide the API interface of Chat-GPT to perform statement translation processing on the dialogue statement.

[0130] In step S430, according to the prompt information, the programming interface of the deep learning model is called to perform sensitive word filtering on the dialogue statement to obtain the target statement.

[0131] Since the needs of users are diverse, the dialogue statements input by users may contain sensitive words such as politics and violence. These sensitive words need to be correctly identified and sensitive word filtering is performed.

[0132] In addition to the pre-set sensitive words, some custom words can also be added as sensitive words according to the actual situation requirements.

[0133] In the present exemplary embodiment, during the development process of developers, the method of writing prompt information can be used to guide the call of the programming interface of the deep learning model to perform model preprocessing on the dialogue statement, which can guide the deep learning model to accurately and reliably learn the corresponding tasks during the development stage, ensuring the accuracy of subsequent statement processing by users, and also expanding the inclusiveness of users from different countries or language habits for dialogue statement input, further optimizing the user experience.

[0134] In step S120, an intent filling template is obtained, and the language model is used to perform semantic recognition processing on the target statement to obtain semantic information.

[0135] In the exemplary embodiment of the present disclosure, an intent filling template can also be further obtained.

[0136] In an alternative embodiment, the intent filling template includes: dialogue questions, dialogue demands, and dialogue entities.

[0137] Among them, the dialogue problem can be the problem reflected by the user during the communication with the customer service regarding the sold goods or other process-related issues; the dialogue requirement can be the purpose that the user wants to achieve through the reflected problem during the communication with the customer service; the dialogue entity can be the specific content uniquely representing the problem reflected by the user, such as the order number, or the user ID (Identity document) etc.

[0138] Further, after using the deep learning model to perform model preprocessing on the dialogue statement to obtain the target statement, the language model can be further used to perform semantic recognition processing on the target statement.

[0139] In an alternative embodiment, the target statement is input into the fine-tuned language model so that the fine-tuned language model performs semantic recognition processing on the target statement to obtain semantic information.

[0140] Before inputting the target statement into the fine-tuned language model, the model can be optimized first.

[0141] In an alternative embodiment, Figure 5 shows a flowchart of a method for model optimization, as Figure 5 shown, the method can at least include the following steps: In step S510, the original sample data is obtained, and the original sample data is preprocessed to obtain the target sample data.

[0142] The original sample data can be the communication dialogues between users and customer service staff in various countries collected in the e-commerce scenario.

[0143] In order to construct a high-quality private user dataset and protect user privacy, after obtaining the original sample data, the original sample data can be preprocessed to complete the processing and processing of the original sample data.

[0144] In an alternative embodiment, Figure 6 shows a flowchart of a method for sample preprocessing, as Figure 6 shown, the method can at least include the following steps: In step S610, the original sample data is cleaned to obtain the first sample data.

[0145] Data cleaning is an important task in data analysis. It refers to cleaning, correcting, formatting, and organizing the original sample data so that it can be converted into data available for analysis.

[0146] For the original sample data with missing characters, misspelled characters, and sensitive information, the original data can be cleaned by means of deletion, induction, interpolation, or extrapolation to obtain the first sample data.

[0147] In step S620, the first sample data is subjected to statement preprocessing to obtain the second sample data.

[0148] In an alternative embodiment, Figure 7 A flowchart showing the method of statement preprocessing is shown, as Figure 7 shown, the method may at least include the following steps: In step S710, the first sample data is subjected to model preprocessing to obtain the third sample data.

[0149] Similar to the model preprocessing method for dialogue statements, it is also possible to use a written prompt message to guide the invocation of the API interface of deep learning models such as Chat-GPT to perform model preprocessing on the original sample data to obtain the third sample data, and the specific process will not be elaborated here.

[0150] In step S720, the third sample data is subjected to label division processing according to the intent filling template to obtain the second sample data.

[0151] Since the intent filling template includes three parts: dialogue question, dialogue requirement, and dialogue entity, after obtaining the third sample data, the relevance of the third sample data to the three parts of dialogue question, dialogue requirement, and dialogue entity can be divided to label the third sample data with "dialogue question, dialogue requirement, or dialogue entity" to obtain the corresponding second sample data.

[0152] Specifically, when determining the relevance of the third sample data to the three parts of dialogue question, dialogue requirement, and dialogue entity, the semantic similarity between each pair can be calculated. Only when the semantic similarity between the third sample data and the dialogue question, dialogue requirement, or dialogue entity is greater than the corresponding threshold, for example, 70%, can the label of the third sample data be determined accordingly. In addition, other means of label division processing can also be selected according to the actual situation, and this exemplary embodiment does not make special limitations on this.

[0153] To ensure the accuracy of the second sample data, it is also possible to manually check and correct the second sample data after the label division processing.

[0154] In step S630, the second sample data is subjected to data augmentation processing to obtain the target sample data.

[0155] After obtaining the second sample data, the API interface of Chat-GPT can be called to perform data augmentation and expansion on the second sample data with the same label to achieve the effect of different texts but the same meaning and label, and increase the sample size of the original sample data, and finally form a large-scale and high-quality target sample data.

[0156] In this exemplary embodiment, the quality of the original sample data is optimized and the quantity is augmented through the process of sample preprocessing, which optimizes the performance of the dataset for fine-tuning the language model in terms of both quantity and quality. The optimization effect of the language model is ensured by using high-quality and large-scale sample data, and the accuracy of semantic recognition processing of the language model is further improved.

[0157] In step S520, based on the target sample data, the language model before fine-tuning is optimized using the supervised fine-tuning technique to obtain the fine-tuned language model.

[0158] After obtaining the target sample data, the target sample data can be applied to the model optimization process of the language model.

[0159] Among them, the language model can be LLaMA (Large Language Model Meta A, a large language model).

[0160] LLaMA is a basic language model set developed by the research team, aiming to provide pre-trained models with broad language understanding capabilities. By using the LLaMA model, researchers and developers can build more advanced natural language processing systems.

[0161] When using the LLaMA model for natural language processing tasks, text can be input into the model, and the model's understanding of the text and the generated results can be obtained. These results can be used for various tasks, such as text classification, named entity recognition, sentiment analysis, etc.

[0162] The LLaMA model architecture includes the following multiple contents:

[0163] Based on the Transformer architecture: The LLaMA model adopts the Transformer architecture, which is an architecture in the field of natural language processing. The Transformer architecture realizes the encoding and decoding of the input text through the self-attention mechanism and the feed-forward neural network layer.

[0164] Parameter range: The parameter range of the LLaMA model ranges from 7B to 65B, which makes it a very large and powerful language model set. By training on trillions of tokens, the LLaMA model can learn rich language knowledge and semantic understanding capabilities.

[0165] Pre-training data: The training dataset of the LLaMA model uses publicly available datasets. The size of the model's training dataset is 1.4T tokens (symbols used to represent words or phrases).

[0166] On this basis, the language model can use the LLaMA2 model.

[0167] LLaMA2 is the next-generation open-source language model, trained on a 2-trillion-token dataset, with the context length extended from 2048 in LLaMA to 4096, enabling it to understand and generate longer texts. It is a series of pre-trained and fine-tuned models with parameter counts ranging from 7 billion to 70 billion. LLaMA2 has two major advantages that set it apart from other open-source LLMs.

[0168] Specifically, LLaMA 2 can be used freely for research and commercial purposes, and LLaMA 2 has a range of different models.

[0169] LLaMA 2 has a commercial license that allows anyone to integrate it into their products and services. This also means that LLaMA 2 can be used for a variety of purposes, such as building chatbots, generating content, creating voice assistants, etc. LLaMA 2 can also be customized and fine-tuned for specific domains and tasks, such as healthcare, education, finance, etc.

[0170] Another advantage of LLaMA 2 is that it offers a range of models with different sizes and capabilities. Depending on the user's needs and resources, the following models can be selected:

[0171] LLaMA-7B: The smallest model, with 7 billion parameters, suitable for resource-constrained devices and applications.

[0172] LLaMA-14B: A medium-sized model with 14 billion parameters, suitable for general-purpose applications and tasks.

[0173] LLaMA-28B: A large model with 28 billion parameters, suitable for high-performance applications and tasks.

[0174] LLaMA-56B: A very large model with 56 billion parameters, suitable for advanced applications and tasks that require more complexity and diversity.

[0175] LLaMA-70B: The largest model, with 70 billion parameters, suitable for state-of-the-art applications and tasks that require the highest quality and performance.

[0176] All these models are pre-trained on 2 trillion tokens of online data and have a context window of 4096 tokens. Additionally, Meta provides a fine-tuned model called LLaMA-2-chat, which is optimized for conversational applications. LLaMA-2-chat is trained on over 1 million human-annotated texts and can generate smooth and relatively accurate responses.

[0177] Specifically, the model tuning process of language models such as LLaMA 2 can be achieved by using supervised fine-tuning technology to obtain a fine-tuned language model.

[0178] SFT (Supervised Fine-Tuning) means pre-training a neural network model, i.e., the source model, on a source dataset. Then, a new neural network model, i.e., the target model, is created. The target model replicates all the model designs and parameters of the source model except for the output layer. These model parameters contain the knowledge learned from the source dataset, and this knowledge is also applicable to the target dataset. The output layer of the source model is closely related to the labels of the source dataset, so it is not adopted in the target model. During fine-tuning, an output layer with an output size equal to the number of classes in the target dataset is added to the target model, and the model parameters of this layer are randomly initialized. When training the target model on the target dataset, it will be trained from scratch to the output layer, and the parameters of the remaining layers are fine-tuned based on the parameters of the source model.

[0179] Specifically, supervised fine-tuning includes the following steps:

[0180] Pre-training: Train a deep learning model on a large-scale dataset, for example, using self-supervised learning or unsupervised learning algorithms for pre-training.

[0181] Fine-tuning: Fine-tune the pre-trained model using the training set of the target task. Usually, only some layers in the pre-trained model are fine-tuned, such as only the last few layers or some intermediate layers of the model. During the fine-tuning process, the model is optimized through the backpropagation algorithm to make the model perform better on the target task.

[0182] Evaluation: Evaluate the fine-tuned model using the test set of the target task to obtain the performance metrics of the model on the target task.

[0183] In this exemplary embodiment, using the supervised fine-tuning technology to perform model tuning on the language model avoids training the model from scratch, speeds up the training process of the language model, and improves the performance of the language model in semantic recognition tasks.

[0184] In an alternative embodiment, Figure 8 shows a schematic flowchart of another method for model tuning, as Figure 8 shown, this method may at least include the following steps: In step S810, obtain the original sample data and perform sample preprocessing on the original sample data to obtain the target sample data.

[0185] The original sample data can be the communication conversations between users and customer service staff in various countries collected in an e-commerce scenario.

[0186] To construct a high-quality private user dataset and protect user privacy, after obtaining the original sample data, the original sample data can be preprocessed to complete the processing and transformation of the original sample data to obtain the target sample data.

[0187] Specifically, the method for preprocessing the original sample data is the same as the Figure 6 method for sample preprocessing shown, and will not be elaborated here.

[0188] In step S820, based on the target sample data, the language model before fine-tuning is optimized using reward modeling technology to obtain a fine-tuned language model.

[0189] After obtaining the target sample data, the target sample data can be applied to the model optimization process of the language model.

[0190] Among them, the language model can be the LLaMA model, or the LLaMA 2 model can be further used. This exemplary embodiment does not make special limitations on this.

[0191] Specifically, reward modeling technology can be used to optimize the language models such as LLaMA 2 to obtain a fine-tuned language model.

[0192] The reward model inputs a text sequence, and the model gives a reward value that conforms to human preferences. The training data for constructing the reward model is generally the same data generated by different language models, and then manually scored.

[0193] Therefore, the purpose of the reward model is to simulate human scoring of texts. There are many available strategies for constructing the reward model. For example, the most direct prediction annotation can output a score or a boolean value according to good or bad; or it is the ranking of prediction results, that is, for the two results corresponding to each input text, use the model to predict which human annotation score is higher.

[0194] In this exemplary embodiment, using reward modeling technology to optimize the language model avoids training the model from scratch and enables the language model to have better task performance in semantic recognition tasks in a short time.

[0195] In an alternative embodiment, Figure 9 shows a schematic flowchart of another method for model optimization processing, as Figure 9 shown, this method can at least include the following steps: In step S910, obtain the original sample data, and preprocess the original sample data to obtain the target sample data.

[0196] The original sample data can be the communication conversations between users and customer service staff in various countries collected in the e-commerce scenario.

[0197] To construct a high-quality private user dataset and protect user privacy, after obtaining the original sample data, the original sample data can be preprocessed to complete the processing and transformation of the original sample data to obtain the target sample data.

[0198] Specifically, the method for preprocessing the original sample data is the same as Figure 6 the sample preprocessing method shown, which will not be elaborated here.

[0199] In step S920, based on the target sample data, the language model before fine-tuning is optimized using human feedback reinforcement learning technology to obtain a fine-tuned language model.

[0200] After obtaining the target sample data, the target sample data can be applied to the model optimization process of the large language model.

[0201] Among them, the language model can be the LLaMA model, or the LLaMA 2 model can be further used. This exemplary embodiment does not make special limitations on this.

[0202] Specifically, human feedback reinforcement learning technology can be used to implement the model optimization process of language models such as LLaMA 2 to obtain a fine-tuned language model.

[0203] RLHF (Reinforcement learning with human feedback, human feedback reinforcement learning) optimizes the language model according to human feedback in the way of reinforcement learning.

[0204] Human feedback reinforcement learning is an advanced AI system training method that combines reinforcement learning with human feedback. It is a method of incorporating the wisdom and experience of human trainers into the model training process to create a more robust learning process. This technology involves using human feedback to create a reward signal and then improving the behavior of the model through reinforcement learning.

[0205] Reinforcement learning, simply put, is a process. In this process, the AI agent learns to make decisions through interaction with the environment and feedback obtained in the form of rewards or punishments. The goal of the agent is to maximize the cumulative reward over time. RLHF enhances this process by replacing or supplementing the predefined reward function with human-generated feedback, thereby allowing the model to better capture complex human preferences and understanding.

[0206] The process of RLHF can be divided into the following steps:

[0207] Initial model training: At the beginning, the AI model is trained using supervised learning, and human trainers provide labeled examples of correct behaviors. The model learns to predict the correct actions or outputs based on the given inputs.

[0208] Collecting human feedback: After the initial model is trained, human trainers provide feedback on the model's performance. They rank the outputs or behaviors of different models according to quality or correctness. This feedback is used to create a reward signal for reinforcement learning.

[0209] Reinforcement learning: Then, the model is fine-tuned using the Proximal Policy Optimization (PPO) algorithm or a similar algorithm, which incorporates the reward signal generated by humans. The model continuously improves its performance by learning from the feedback provided by human trainers.

[0210] Iterative process: The process of collecting human feedback and improving the model through reinforcement learning is repeated, which continuously improves the performance of the model.

[0211] In this exemplary embodiment, using the human feedback reinforcement learning technique to perform model tuning on the language model before fine-tuning enables the language model to obtain feedback signals through interaction with humans to optimize the generated text without the need for artificial specification of the reward function. This allows the language model to better capture human preferences and understanding and provide more natural and accurate text outputs, with better performance in semantic recognition tasks.

[0212] After performing model tuning on the language model through supervised fine-tuning techniques, reward modeling techniques, or human feedback reinforcement learning techniques, a fine-tuned language model can be obtained.

[0213] Therefore, the target statement obtained by preprocessing the model can be input into the fine-tuned language model, so that the fine-tuned language model performs semantic recognition processing on the target statement to obtain semantic information, further achieving the effect of accurately classifying the user's intention.

[0214] In step S130, intention filling processing is performed on the semantic information according to the intention filling template to obtain the target intention.

[0215] In the exemplary embodiment of the present disclosure, after obtaining the intention filling template and the semantic information obtained by semantic recognition processing, intention filling processing can be performed on the semantic information according to the intention filling template.

[0216] In an alternative embodiment, Figure 10 shows a flowchart of the method for intention filling processing, as Figure 10As shown, the method may at least include the following steps: In step S1010, perform intent filling processing on semantic information according to an intent filling template to obtain a category intent.

[0217] Since the intent filling template includes three parts: a dialogue question, a dialogue request, and a dialogue entity, after obtaining the semantic information, the semantic information can be filled into the corresponding part to obtain a category intent.

[0218] For example, when the semantic information is "return goods", the semantic information can be filled into the "dialogue request" part, and the category intent is determined to be "the dialogue request is to return goods".

[0219] In step S1020, determine the target intent according to the filling situation of the category intent in the intent filling template.

[0220] Since the intent filling template includes three parts: a dialogue question, a dialogue request, and a dialogue entity, it is necessary to judge that the category intent has been filled in all three parts before further determining the user's target intent.

[0221] In an optional embodiment, Figure 11 shows a flowchart of a method for determining a target intent, as Figure 11 shown, the method may at least include the following steps: In step S1110, when the category intent is filled in the intent filling template, determine the category intent in the intent filling template as the target intent.

[0222] For example, when the category intent corresponding to the dialogue question is "the dialogue question is that the quality of the goods I bought is poor", the category intent corresponding to the dialogue request is "the dialogue request is to return goods", and the category intent corresponding to the dialogue entity is "the dialogue entity is order number 0001", it can be judged that none of the three category intents is empty, so the category intent filled into the intent filling template can be determined as the target intent.

[0223] In step S1120, when the category intent is not filled in the intent filling template, obtain the next statement of the dialogue statement to determine the target intent according to the dialogue statement and the next statement.

[0224] For example, when the category intention corresponding to the conversation question is "The quality of the product I bought is poor", the category intention corresponding to the conversation request is "The conversation request is to return the product", but the category intention corresponding to the conversation entity is empty, or when the category intention corresponding to the conversation question is "The quality of the product I bought is poor", and both the category intention corresponding to the conversation request and the category intention corresponding to the conversation entity are empty, it is determined that the current category intention does not fill the three parts in the intention filling template, that is, the user intention at this time cannot meet the subsequent business requirements, or some content is missing. Therefore, the next round of conversation interaction with the user can be further initiated to enable the user to continue to input the next statement, and the next statement is continued to be subjected to processes such as model preprocessing, semantic recognition processing, and intention filling processing to achieve intention recognition until the three parts of the intention filling template are filled with each category intention to obtain the target intention.

[0225] In this exemplary embodiment, determining the corresponding target intention according to the filling situation in the intention filling template improves the accuracy of intention recognition, and also ensures the completeness of the necessary information in the e-commerce scenario, and can provide data guarantee and theoretical basis for the accurate execution of the business logic.

[0226] After determining the user's target intention, the customer service can continue to execute the subsequent business logic. In the e-commerce scenario, the business logic may include operations such as returning goods, exchanging goods, and querying orders.

[0227] The following makes a detailed description of the intention recognition method in the embodiments of the present disclosure in combination with an application scenario.

[0228] Figure 12 The flowchart of the intention recognition method in the application scenario is shown, as Figure 12 shown, in step S1210, the conversation starts.

[0229] In the e-commerce scenario, when the user enters the customer service conversation platform, the user can send a conversation statement to the merchant to start a conversation with the intelligent customer service; or during the development process by developers, a conversation statement can also be input for subsequent development steps.

[0230] In step S1220, statement correction.

[0231] After obtaining the conversation statement, the deep learning model can be used to perform model preprocessing on the conversation statement.

[0232] Among them, the deep learning model can be a generative pre-trained sequence model based on the attention mechanism.

[0233] Chat-GPT is a version that is fine-tuned and refined based on the GPT model to make it more suitable for dialogue generation tasks. Therefore, Chat-GPT can be further utilized to implement the process of model preprocessing.

[0234] First, call the programming interface of the deep learning model to perform sentence correction on the dialogue sentences to obtain the target sentences.

[0235] When the deep learning model is Chat-GPT, the API interface of Chat-GPT can be called, and this API interface is used to perform sentence correction on the dialogue sentences to obtain the corresponding target sentences.

[0236] When developers input dialogue sentences during the development process, in order to be able to call the API interface of Chat-GPT to perform the corresponding model preprocessing operations, a Prompt can be written using prompt engineering to guide deep models such as Chat-GPT to achieve.

[0237] In step S1230, sentence translation.

[0238] After the input dialogue sentences are corrected correctly, considering the problem of foreign users using multiple languages, a Prompt can continue to be used to guide the API interface of Chat-GPT to perform sentence translation processing on the corrected dialogue sentences.

[0239] In step S1240, sensitive word filtering.

[0240] Since users' needs are diverse, the dialogue sentences input by users may contain sensitive words such as politics and violence. These sensitive words need to be correctly identified and sensitive word filtering is performed.

[0241] In addition to the pre-set sensitive words, some custom words can also be added as sensitive words according to the actual situation requirements.

[0242] In step S1250, semantic recognition.

[0243] After using the deep learning model to perform model preprocessing on the dialogue sentences to obtain the target sentences, the language model can be further utilized to perform semantic recognition processing on the target sentences.

[0244] Input the target sentences into the fine-tuned language model so that the fine-tuned language model performs semantic recognition processing on the target sentences to obtain semantic information.

[0245] Before inputting the target sentences into the fine-tuned language model, the language model before fine-tuning can be first optimized.

[0246] Figure 13 The flowchart shows the method of sample preprocessing in an application scenario, as Figure 17 shown, in step S1310, the native dataset.

[0247] Among them, the language model can be the LLaMA model, or the LLaMA 2 model can be further used. This exemplary embodiment does not make special limitations on this.

[0248] Specifically, the LLaMA 2 model is an open-source and commercially available large model algorithm. After deploying language models such as LLaMA 2 locally, the native dataset can be obtained first.

[0249] The native dataset can be the communication conversations between users and customer service staff in various countries collected in the e-commerce scenario and used as the original sample data.

[0250] In step S1320, data cleaning.

[0251] In order to construct a high-quality private user dataset and protect user privacy, after obtaining the original sample data, the original sample data can be preprocessed to complete the processing and processing of the original sample data.

[0252] For the original sample data with missing characters, misspelled characters, and sensitive information, the original data can be processed by data cleaning methods such as deletion, induction, interpolation, or extrapolation to obtain the first sample data.

[0253] In step S1330, GPT processing.

[0254] The first sample data is preprocessed by a model to obtain the third sample data.

[0255] Similar to the model preprocessing method for dialogue statements, the API interface of deep learning models such as Chat-GPT can also be called by writing prompt information to preprocess the original sample data to obtain the third sample data. The specific process of model preprocessing will not be elaborated here.

[0256] Furthermore, the third sample data is processed by label division according to the intention filling template to obtain the second sample data.

[0257] Since the intention filling template includes three parts: dialogue questions, dialogue demands, and dialogue entities, after obtaining the third sample data, the relevance of the third sample data to the three parts of dialogue questions, dialogue demands, and dialogue entities can be divided to label the third sample data with "dialogue questions, dialogue demands, or dialogue entities" to obtain the corresponding second sample data.

[0258] Specifically, when determining the relevance of the third sample data to the three parts of the dialogue question, dialogue appeal, and dialogue entity, the semantic similarity between each pair can be calculated. Only when the semantic similarity between the third sample data and the dialogue question, dialogue appeal, or dialogue entity is greater than the corresponding threshold, for example, 70%, can the label of the third sample data be determined accordingly. In addition, other means of label division processing can also be selected according to the actual situation, and this exemplary embodiment does not make special limitations on this.

[0259] In step S1340, manual review.

[0260] To ensure the accuracy of the second sample data, after the label division processing, manual inspection and correction of the second sample data can also be performed.

[0261] In step S1350, GPT data augmentation.

[0262] After obtaining the second sample data, the API interface of Chat-GPT can be called to perform data augmentation and expansion on the second sample data with the same label, so as to achieve the effect of different texts but the same meaning and label, and increase the sample size of the original sample data. Finally, a large-scale and high-quality target sample data is formed for training language models such as LLaMA 2.

[0263] In step S1360, data set.

[0264] After data augmentation and expansion of the second sample data, a data set, that is, the target sample data, is obtained. After the target sample data is constructed, the target sample data can be applied in the pre-training and tuning processes of the language model.

[0265] Figure 14 Shows a schematic flowchart of the method for model tuning processing in the application scenario, as Figure 14 shown, in step S1410, data set.

[0266] In step S1420, self-supervised learning.

[0267] Self-Supervised Learning, also known as self-supervised learning, generally machine learning is divided into supervised learning, unsupervised learning, and reinforcement learning. And Self-Supervised Learning is a type of unsupervised learning, mainly hoping to learn a general feature representation for downstream tasks, and the main way is to supervise itself.

[0268] In step S1430, LLaMA 2.

[0269] In step S1440, supervised fine-tuning.

[0270] Specifically, the model tuning process of language models such as LLaMA 2 is achieved by using supervised fine-tuning technology to obtain a fine-tuned language model.

[0271] SFT means pre-training a neural network model, i.e., the source model, on the source dataset. Then a new neural network model, i.e., the target model, is created.

[0272] The target model replicates all the model designs and their parameters of the source model except for the output layer. These model parameters contain the knowledge learned on the source dataset, and this knowledge also applies to the target dataset. The output layer of the source model is closely related to the labels of the source dataset, so it is not adopted in the target model.

[0273] During fine-tuning, an output layer with an output size equal to the number of classes in the target dataset is added to the target model, and the model parameters of this layer are randomly initialized. When training the target model on the target dataset, it will be trained from scratch to the output layer, and the parameters of the remaining layers are fine-tuned based on the parameters of the source model.

[0274] In addition, other common methods for the model tuning process of language models include reward modeling technology and human feedback reinforcement learning technology, etc.

[0275] Specifically, the model tuning process of the language model before fine-tuning is carried out using reward modeling technology to obtain a fine-tuned language model.

[0276] The reward model inputs a text sequence, and the model gives a reward value that conforms to human preferences. The training data for constructing the reward model is generally the same data generated by different language models, and then manually scored.

[0277] Therefore, the purpose of the reward model is to simulate human scoring of texts. There are many available strategies for constructing the reward model. For example, the most direct prediction annotation can output a score or a boolean value according to good or bad; or it is the ranking of prediction results, that is, for the two results corresponding to each input text, use the model to predict which score of human annotation is higher.

[0278] The model tuning process of the language model before fine-tuning is carried out using human feedback reinforcement learning technology to obtain a fine-tuned language model.

[0279] Human feedback reinforcement learning is an advanced method for training AI systems. It combines reinforcement learning with human feedback. It is a method of creating a more robust learning process by incorporating the wisdom and experience of human trainers into the model training process. This technology involves using human feedback to create a reward signal, and then improving the behavior of the model through reinforcement learning.

[0280] Reinforcement learning is a process. In this process, an AI agent learns to make decisions through interactions with the environment and feedback received in the form of rewards or punishments. The goal of the agent is to maximize the cumulative reward over time. RLHF enhances this process by replacing or supplementing the predefined reward function with human-generated feedback, thereby allowing the model to better capture complex human preferences and understanding.

[0281] By means of techniques such as supervised fine-tuning technology, reward modeling technology, and human feedback reinforcement learning technology, fine-tuning and training language models such as LLaMA 2 can achieve the effect of accurately classifying user intentions.

[0282] In step S1450, user & customer service feedback.

[0283] In a formal usage environment, since new users continuously generate new conversation data, the data generated in the environment and the correct intentions can be used as a sample data set for enhancing the language model, and the language model can be continuously updated and enhanced.

[0284] Among them, the data generated in the environment and the corresponding correct intentions can be correctly identified by the algorithm itself or obtained after manual inspection and label correction. This exemplary embodiment does not make special limitations on this.

[0285] Therefore, the target statement obtained by preprocessing the model can be input into the fine-tuned language model, so that the fine-tuned language model performs semantic recognition processing on the target statement to obtain semantic information, and further achieves the effect of accurately classifying user intentions.

[0286] In step S1260, intention filling.

[0287] Furthermore, obtain an intention filling template. The intention filling template includes a conversation question, a conversation request, and a conversation entity.

[0288] Among them, the conversation question can be a problem related to the sold goods or other processes reflected by the user during the communication with the customer service; the conversation request can be the purpose that the user wants to achieve through the reflected problem during the communication with the customer service; the conversation entity can be the specific content uniquely representing the problem reflected by the user, such as an order number, or a user ID, etc.

[0289] Perform intention filling processing on the semantic information according to the intention filling template to obtain a category intention.

[0290] Since the intention filling template includes three parts: a conversation question, a conversation request, and a conversation entity, after obtaining the semantic information, the semantic information can be filled into the corresponding part to obtain a category intention.

[0291] For example, when the semantic information is "return goods", this semantic information can be filled into the "dialogue request" part to determine that the category intention is "the dialogue request is to return goods".

[0292] Determine the target intention according to the filling situation of the category intention in the intention filling template.

[0293] Since the intention filling template includes three parts: dialogue question, dialogue request, and dialogue entity, it is necessary to judge that the category intentions of all three parts have been filled in order to further determine the user's target intention.

[0294] In step S1270, output the user intention.

[0295] When the category intention in the intention filling template is filled, determine the category intention in the intention filling template as the target intention.

[0296] For example, when the category intention corresponding to the dialogue question is "the dialogue question is that the quality of the goods I bought is poor", the category intention corresponding to the dialogue request is "the dialogue request is to return goods", and the category intention corresponding to the dialogue entity is "the dialogue entity is order number 0001", it can be judged that none of the three category intentions are empty. Therefore, the category intention filled into the intention filling template can be determined as the target intention.

[0297] In step S1280, intention guidance.

[0298] When the category intention in the intention filling template is not filled, obtain the next statement of the dialogue statement to determine the target intention according to the dialogue statement and the next statement.

[0299] For example, when the category intention corresponding to the dialogue question is "the dialogue question is that the quality of the goods I bought is poor", the category intention corresponding to the dialogue request is "the dialogue request is to return goods", but the category intention corresponding to the dialogue entity is empty, or when the category intention corresponding to the dialogue question is "the dialogue question is that the quality of the goods I bought is poor", and the category intentions corresponding to the dialogue request and the dialogue entity are both empty, it is determined that the current category intention does not fill the three parts in the intention filling template, that is, the user intention at this time cannot meet the subsequent business requirements, or some content is missing. Therefore, the next round of dialogue interaction with the user can be further initiated to enable the user to continue to input the next statement, and then continue to perform processes such as model preprocessing, semantic recognition processing, and intention filling processing on the next statement to achieve intention recognition until the category intentions of each part fill the three parts of the intention filling template to obtain the target intention.

[0300] In an exemplary embodiment of the present disclosure, a deep learning model is used to preprocess the dialogue statements, constructing a high-quality dataset, solving the problem of low data quality caused by misspelled words and missing words in the dialogue statements, as well as the problem of high translation costs for multi-language statements. Further, it provides data support for providing a general intention recognition method, improving the accuracy of intention recognition from a data perspective, and also being able to enhance the stability of the method by filtering sensitive words or content. Further, a language model is used to perform semantic recognition processing on the target statement, leveraging the powerful reasoning ability of the language model and the learning advantages of more hyperparameter factors for the intention recognition task, improving the accuracy of intention recognition from the model itself. In addition, through the language model, it is also possible to effectively expand and enhance the dataset, forming a large-scale and high-quality dataset for the intention recognition task, and being able to deploy the language model locally to ensure the privacy and security of the data, which is of great significance in practical applications. Furthermore, the semantic information is filled into the corresponding intention filling template to obtain the user's target intention, providing a basis and foundation for the timely execution of the business logic, providing a general method for intention recognition of multi-language dialogue statements, optimizing the accuracy of intention recognition from multiple perspectives, greatly optimizing the user experience, and improving the user return rate to a certain extent.

[0301] In addition, in an exemplary embodiment of the present disclosure, an intention recognition device is also provided. Figure 15 The structural schematic diagram of the intention recognition device is shown, as Figure 15 shown, the intention recognition device 1500 may include:

[0302] A statement acquisition module 1510, configured to acquire dialogue statements and perform model preprocessing on the dialogue statements using a deep learning model to obtain a target statement;

[0303] A semantic recognition module 1520, configured to acquire an intention filling template and perform semantic recognition processing on the target statement using a language model to obtain semantic information;

[0304] An intention filling module 1530, configured to perform intention filling processing on the semantic information according to the intention filling template to obtain a target intention.

[0305] In some embodiments of the present disclosure, the statement acquisition module 1510 includes:

[0306] A first correction sub-module, configured to call the programming interface of the deep learning model to perform statement correction processing on the dialogue statements to obtain a target statement; and / or

[0307] A first translation sub-module, configured to call the programming interface of the deep learning model to perform statement translation processing on the dialogue statements to obtain a target statement; and / or

[0308] The first filtering sub-module is configured to call the programming interface of the deep learning model to filter sensitive words from the dialogue statement to obtain a target statement.

[0309] In some embodiments of the present disclosure, the statement acquisition module 1510 includes:

[0310] A prompt acquisition sub-module configured to acquire prompt information;

[0311] A model processing sub-module configured to call the programming interface of the deep learning model according to the prompt information to perform model preprocessing on the dialogue statement to obtain a target statement.

[0312] In some embodiments of the present disclosure, the model processing sub-module includes:

[0313] A second correction unit configured to call the programming interface of the deep learning model according to the prompt information to perform statement correction processing on the dialogue statement to obtain a target statement; and / or

[0314] A second translation unit configured to call the programming interface of the deep learning model according to the prompt information to perform statement translation processing on the dialogue statement to obtain a target statement; and / or

[0315] A second filtering unit configured to call the programming interface of the deep learning model according to the prompt information to filter sensitive words from the dialogue statement to obtain a target statement.

[0316] In some embodiments of the present disclosure, in the statement acquisition module 1510, the deep learning model includes: a generative pre-trained sequence model based on an attention mechanism.

[0317] In some embodiments of the present disclosure, the semantic recognition module 1520 includes:

[0318] A model recognition sub-module configured to input the target statement into a fine-tuned language model, so that the fine-tuned language model performs semantic recognition processing on the target statement to obtain semantic information.

[0319] In some embodiments of the present disclosure, the intent recognition device 1500 further includes:

[0320] A first acquisition module configured to acquire original sample data and perform sample preprocessing on the original sample data to obtain target sample data;

[0321] A supervised fine-tuning module configured to perform model tuning processing on a pre-fine-tuned language model based on the target sample data by using supervised fine-tuning technology to obtain a fine-tuned language model.

[0322] In some embodiments of the present disclosure, the intent recognition device 1500 further includes:

[0323] A second acquisition module, configured to acquire original sample data and perform sample preprocessing on the original sample data to obtain target sample data;

[0324] A reward modeling module, configured to perform model tuning on the pre-fine-tuning language model based on the target sample data using reward modeling technology to obtain a fine-tuned language model.

[0325] In some embodiments of the present disclosure, the intent recognition device 1500 further includes:

[0326] A third acquisition module, configured to acquire original sample data and perform sample preprocessing on the original sample data to obtain target sample data;

[0327] A reinforcement learning module, configured to perform model tuning on the pre-fine-tuning language model based on the target sample data using human feedback reinforcement learning technology to obtain a fine-tuned language model.

[0328] In some embodiments of the present disclosure, the first acquisition module, or the second acquisition module, or the third acquisition module includes

[0329] A data cleaning sub-module, configured to perform data cleaning on the original sample data to obtain first sample data;

[0330] A statement processing sub-module, configured to perform statement preprocessing on the first sample data to obtain second sample data;

[0331] A data augmentation sub-module, configured to perform data augmentation on the second sample data to obtain target sample data.

[0332] In some embodiments of the present disclosure, the intent filling template includes: dialogue questions, dialogue demands, and dialogue entities.

[0333] In some embodiments of the present disclosure, the statement processing sub-module includes:

[0334] A sample processing unit, configured to perform model preprocessing on the first sample data to obtain third sample data;

[0335] A label partitioning unit, configured to perform label partitioning on the third sample data according to the intent filling template to obtain second sample data.

[0336] In some embodiments of the present disclosure, the label partitioning unit includes:

[0337] A filling processing subunit, configured to perform intent filling processing on the semantic information according to the intent filling template to obtain a category intent;

[0338] An intent filling subunit, configured to determine a target intent according to the filling situation of the category intent in the intent filling template.

[0339] In some embodiments of the present disclosure, the intent filling subunit includes:

[0340] A filling completed component, configured to determine the category intent in the intent filling template as the target intent when the intent filling template fills the category intent; or

[0341] A filling unfinished component, configured to obtain the next statement of the dialogue statement when the intent filling template does not fill the category intent, so as to determine the target intent according to the dialogue statement and the next statement.

[0342] Regarding the device in the above embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0343] The present disclosure also provides a computer-readable storage medium, on which computer program instructions are stored, and when the program instructions are executed by a processor, the steps of the intent recognition method provided by the present disclosure are implemented.

[0344] Figure 16 It is a block diagram of another intent recognition device 1600 shown according to an exemplary embodiment. For example, the device 1600 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0345] Refer to Figure 16 , the device 1600 may include one or more of the following components: a processing component 1602, a memory 1604, a power supply component 1606, a multimedia component 1608, an audio component 1610, an input / output interface 1612, a sensor component 1614, and a communication component 1616.

[0346] The processing component 1602 generally controls the overall operation of the device 1600, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations. The processing component 1602 may include one or more processors 1620 to execute instructions to complete all or part of the steps of the above methods. In addition, the processing component 1602 may include one or more modules to facilitate the interaction between the processing component 1602 and other components. For example, the processing component 1602 may include a multimedia module to facilitate the interaction between the multimedia component 1608 and the processing component 1602.

[0347] The memory 1604 is configured to store various types of data to support the operation of the device 1600. Examples of such data include instructions for any application or method operating on the device 1600, contact data, phone book data, messages, pictures, videos, etc. The memory 1604 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0348] The power component 1606 provides power to various components of the device 1600. The power component 1606 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 1600.

[0349] The multimedia component 1608 includes a screen that provides an output interface between the device 1600 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of the touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation. In some embodiments, the multimedia component 1608 includes a front camera and / or a rear camera. When the device 1600 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each of the front camera and the rear camera may be a fixed optical lens system or have a focal length and optical zoom capabilities.

[0350] The audio component 1610 is configured to output and / or input audio signals. For example, the audio component 1610 includes a microphone (MIC) that is configured to receive external audio signals when the device 1600 is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signals can be further stored in the memory 1604 or transmitted via the communication component 1616. In some embodiments, the audio component 1610 further includes a speaker for outputting audio signals.

[0351] The input / output interface 1612 provides an interface between the processing component 1602 and peripheral interface modules, and the peripheral interface modules may be a keyboard, a click wheel, buttons, etc. These buttons may include, but are not limited to: a home button, a volume button, a start button, and a lock button.

[0352] The sensor component 1614 includes one or more sensors for providing an assessment of the status of various aspects of the device 1600. For example, the sensor component 1614 can detect the open / closed state of the device 1600, the relative positioning of components, such as the display and keypad of the device 1600, the sensor component 1614 can also detect a change in the position of the device 1600 or a component of the device 1600, the presence or absence of user contact with the device 1600, the orientation or acceleration / deceleration of the device 1600, and the temperature change of the device 1600. The sensor component 1614 can include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor component 1614 can also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor component 1614 can further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0353] The communication component 1616 is configured to facilitate communication between the device 1600 and other devices in a wired or wireless manner. The device 1600 can access a wireless network based on communication standards, such as WiFi, 2G, or 3G, or a combination thereof. In an exemplary embodiment, the communication component 1616 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 1616 further includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0354] In an exemplary embodiment, the apparatus 1600 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components for performing the above method.

[0355] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 1604 including instructions, and the above instructions can be executed by a processor 1620 of the apparatus 1600 to complete the above method. For example, the non-transitory computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0356] In addition to being an independent electronic device, the above apparatus may also be a part of an independent electronic device. For example, in one embodiment, the apparatus may be an integrated circuit (IC) or a chip. The integrated circuit may be a single IC or a collection of multiple ICs. The chip may include, but is not limited to, the following types: GPU (Graphics Processing Unit), CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), SOC (System on Chip), etc. The above integrated circuit or chip may be used to execute executable instructions (or code) to implement the above intention recognition method. The executable instructions may be stored in the integrated circuit or chip, or obtained from other devices or equipment. For example, the integrated circuit or chip includes a processor, a memory, and an interface for communicating with other devices. The executable instructions may be stored in the memory, and when the executable instructions are executed by the processor, the above intention recognition method is implemented; or, the integrated circuit or chip may receive the executable instructions through the interface and transmit them to the processor for execution to implement the above intention recognition method.

[0357] In another exemplary embodiment, a computer program product is also provided. The computer program product includes a computer program that can be executed by a programmable device. The computer program has a code portion for performing the above-described intent recognition method when executed by the programmable device.

[0358] Figure 17 FIG. 4 is a block diagram of yet another intent recognition device 1700 shown in accordance with an exemplary embodiment. For example, device 1700 may be provided as a server. Referring Figure 17 to FIG. 4, device 1700 includes a processing component 1722, which further includes one or more processors, and memory resources represented by a memory 1732 for storing instructions executable by the processing component 1722, such as application programs. The application programs stored in memory 1732 may include one or more modules each corresponding to a set of instructions. In addition, the processing component 1722 is configured to execute instructions to perform the above-described method...

[0359] Device 1700 may also include a power component 1726 configured to perform power management of device 1700, a wired or wireless network interface 1750 configured to connect device 1700 to a network, and an input / output interface 1758. Device 1700 may operate based on an operating system stored in memory 1732, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM or the like.

[0360] Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include known common knowledge or conventional technical means in the technical field not disclosed by the present disclosure. The specification and embodiments are only to be considered exemplary, and the true scope and spirit of the present disclosure are pointed out by the following claims.

[0361] It should be understood that the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present disclosure is only limited by the appended claims.

Claims

1. An intention recognition method, characterized in that, Including: Obtain a dialogue statement, and perform model preprocessing on the dialogue statement using a deep learning model to obtain a target statement; Obtain an intent filling template, and perform semantic recognition processing on the target statement using a language model to obtain semantic information; Perform intent filling processing on the semantic information according to the intent filling template to obtain a target intent.

2. The intention recognition method according to claim 1, characterized in that, The performing model preprocessing on the dialogue statement using a deep learning model to obtain a target statement includes: Invoking the programming interface of the deep learning model to perform statement correction processing on the dialogue statement to obtain a target statement; and / or Invoking the programming interface of the deep learning model to perform statement translation processing on the dialogue statement to obtain a target statement; and / or Invoking the programming interface of the deep learning model to perform sensitive word filtering on the dialogue statement to obtain a target statement.

3. The intention recognition method according to claim 1, characterized in that, The performing model preprocessing on the dialogue statement using a deep learning model to obtain a target statement includes: Obtain a prompt message; According to the prompt message, invoke the programming interface of the deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement.

4. The intention recognition method according to claim 3, characterized in that, The according to the prompt message, invoking the programming interface of the deep learning model to perform model preprocessing on the dialogue statement to obtain a target statement includes: According to the prompt message, invoke the programming interface of the deep learning model to perform statement correction processing on the dialogue statement to obtain a target statement; and / or According to the prompt message, invoke the programming interface of the deep learning model to perform statement translation processing on the dialogue statement to obtain a target statement; and / or According to the prompt message, invoke the programming interface of the deep learning model to perform sensitive word filtering on the dialogue statement to obtain a target statement.

5. The intention recognition method according to claim 1, characterized in that, The deep learning model includes: a generative pre-trained sequence model based on an attention mechanism.

6. The intention recognition method according to claim 1, characterized in that, The performing semantic recognition processing on the target statement using a language model to obtain semantic information includes: Input the target statement into the fine-tuned language model, so that the fine-tuned language model performs semantic recognition processing on the target statement to obtain semantic information.

7. The intention recognition method according to claim 6, characterized in that, Before the inputting the target statement into the fine-tuned language model, the method further includes: Obtain original sample data, and perform sample preprocessing on the original sample data to obtain target sample data; Based on the target sample data, use the supervised fine-tuning technique to perform model tuning processing on the pre-fine-tuning language model to obtain a fine-tuned language model.

8. The intention recognition method according to claim 6, characterized in that, Before the inputting the target statement into the fine-tuned language model, the method further includes: Obtain original sample data, and perform sample preprocessing on the original sample data to obtain target sample data; Based on the target sample data, use the reward modeling technique to perform model tuning processing on the pre-fine-tuning language model to obtain a fine-tuned language model.

9. The intention recognition method according to claim 6, characterized in that, Before the inputting the target statement into the fine-tuned language model, the method further includes: Obtain original sample data, and perform sample preprocessing on the original sample data to obtain target sample data; Based on the target sample data, use the human feedback reinforcement learning technique to perform model tuning processing on the pre-fine-tuning language model to obtain a fine-tuned language model.

10. The intention recognition method according to any one of claims 7-9, characterized in that, The preprocessing of the original sample data to obtain the target sample data includes: Performing data cleaning on the original sample data to obtain the first sample data; Performing sentence preprocessing on the first sample data to obtain the second sample data; Performing data augmentation on the second sample data to obtain the target sample data.

11. The intention recognition method according to claim 10, characterized in that, The intention filling template includes: dialogue questions, dialogue demands, and dialogue entities.

12. The intention recognition method according to claim 11, characterized in that, The performing sentence preprocessing on the first sample data to obtain the second sample data includes: Performing model preprocessing on the first sample data to obtain the third sample data; Performing label division on the third sample data according to the intention filling template to obtain the second sample data.

13. The intention recognition method according to claim 11, characterized in that, The performing intention filling on the semantic information according to the intention filling template to obtain the target intention includes: Performing intention filling on the semantic information according to the intention filling template to obtain the category intention; Determining the target intention according to the filling situation of the category intention in the intention filling template.

14. The intention recognition method according to claim 13, characterized in that, The determining the target intention according to the filling situation of the category intention in the intention filling template includes: When the intention filling template is filled with the category intention, determining the category intention in the intention filling template as the target intention; or When the intention filling template is not filled with the category intention, obtaining the next sentence of the dialogue sentence to determine the target intention according to the dialogue sentence and the next sentence.

15. An intention recognition device, characterized in that, It includes: A sentence acquisition module configured to acquire a dialogue sentence and perform model preprocessing on the dialogue sentence using a deep learning model to obtain a target sentence; A semantic recognition module configured to acquire an intention filling template and perform semantic recognition on the target sentence using a language model to obtain semantic information; An intention filling module configured to perform intention filling on the semantic information according to the intention filling template to obtain a target intention.

16. A computer-readable storage medium, on which computer program instructions are stored, characterized in that, When the program instruction is executed by a processor, it implements the steps of the method according to any one of claims 1 to 14.

17. An electronic device, characterized in that, It includes: A memory storing a computer program thereon; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1 to 14.

Images